1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file implements a TargetTransformInfo analysis pass specific to the
11 /// X86 target machine. It uses the target's detailed information to provide
12 /// more precise answers to certain TTI queries, while letting the target
13 /// independent and default TTI implementations handle the rest.
15 //===----------------------------------------------------------------------===//
18 #include "X86TargetMachine.h"
19 #include "llvm/ADT/DepthFirstIterator.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/TargetTransformInfo.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Target/CostTable.h"
26 #include "llvm/Target/TargetLowering.h"
29 #define DEBUG_TYPE "x86tti"
31 // Declare the pass initialization routine locally as target-specific passes
32 // don't havve a target-wide initialization entry point, and so we rely on the
33 // pass constructor initialization.
35 void initializeX86TTIPass(PassRegistry &);
39 UsePartialUnrolling("x86-use-partial-unrolling", cl::init(true),
40 cl::desc("Use partial unrolling for some X86 targets"), cl::Hidden);
41 static cl::opt<unsigned>
42 PartialUnrollingThreshold("x86-partial-unrolling-threshold", cl::init(0),
43 cl::desc("Threshold for X86 partial unrolling"), cl::Hidden);
44 static cl::opt<unsigned>
45 PartialUnrollingMaxBranches("x86-partial-max-branches", cl::init(2),
46 cl::desc("Threshold for taken branches in X86 partial unrolling"),
51 class X86TTI final : public ImmutablePass, public TargetTransformInfo {
52 const X86Subtarget *ST;
53 const X86TargetLowering *TLI;
55 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
56 /// are set if the result needs to be inserted and/or extracted from vectors.
57 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
60 X86TTI() : ImmutablePass(ID), ST(nullptr), TLI(nullptr) {
61 llvm_unreachable("This pass cannot be directly constructed");
64 X86TTI(const X86TargetMachine *TM)
65 : ImmutablePass(ID), ST(TM->getSubtargetImpl()),
66 TLI(TM->getTargetLowering()) {
67 initializeX86TTIPass(*PassRegistry::getPassRegistry());
70 void initializePass() override {
74 void getAnalysisUsage(AnalysisUsage &AU) const override {
75 TargetTransformInfo::getAnalysisUsage(AU);
78 /// Pass identification.
81 /// Provide necessary pointer adjustments for the two base classes.
82 void *getAdjustedAnalysisPointer(const void *ID) override {
83 if (ID == &TargetTransformInfo::ID)
84 return (TargetTransformInfo*)this;
88 /// \name Scalar TTI Implementations
90 PopcntSupportKind getPopcntSupport(unsigned TyWidth) const override;
91 void getUnrollingPreferences(Loop *L,
92 UnrollingPreferences &UP) const override;
96 /// \name Vector TTI Implementations
99 unsigned getNumberOfRegisters(bool Vector) const override;
100 unsigned getRegisterBitWidth(bool Vector) const override;
101 unsigned getMaximumUnrollFactor() const override;
102 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
103 OperandValueKind) const override;
104 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
105 int Index, Type *SubTp) const override;
106 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
107 Type *Src) const override;
108 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
109 Type *CondTy) const override;
110 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
111 unsigned Index) const override;
112 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
113 unsigned AddressSpace) const override;
115 unsigned getAddressComputationCost(Type *PtrTy,
116 bool IsComplex) const override;
118 unsigned getReductionCost(unsigned Opcode, Type *Ty,
119 bool IsPairwiseForm) const override;
121 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const override;
123 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
124 Type *Ty) const override;
125 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
126 Type *Ty) const override;
131 } // end anonymous namespace
133 INITIALIZE_AG_PASS(X86TTI, TargetTransformInfo, "x86tti",
134 "X86 Target Transform Info", true, true, false)
138 llvm::createX86TargetTransformInfoPass(const X86TargetMachine *TM) {
139 return new X86TTI(TM);
143 //===----------------------------------------------------------------------===//
147 //===----------------------------------------------------------------------===//
149 X86TTI::PopcntSupportKind X86TTI::getPopcntSupport(unsigned TyWidth) const {
150 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
151 // TODO: Currently the __builtin_popcount() implementation using SSE3
152 // instructions is inefficient. Once the problem is fixed, we should
153 // call ST->hasSSE3() instead of ST->hasPOPCNT().
154 return ST->hasPOPCNT() ? PSK_FastHardware : PSK_Software;
157 void X86TTI::getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const {
158 if (!UsePartialUnrolling)
160 // According to the Intel 64 and IA-32 Architectures Optimization Reference
161 // Manual, Intel Core models and later have a loop stream detector
162 // (and associated uop queue) that can benefit from partial unrolling.
163 // The relevant requirements are:
164 // - The loop must have no more than 4 (8 for Nehalem and later) branches
165 // taken, and none of them may be calls.
166 // - The loop can have no more than 18 (28 for Nehalem and later) uops.
168 // According to the Software Optimization Guide for AMD Family 15h Processors,
169 // models 30h-4fh (Steamroller and later) have a loop predictor and loop
170 // buffer which can benefit from partial unrolling.
171 // The relevant requirements are:
172 // - The loop must have fewer than 16 branches
173 // - The loop must have less than 40 uops in all executed loop branches
175 unsigned MaxBranches, MaxOps;
176 if (PartialUnrollingThreshold.getNumOccurrences() > 0) {
177 MaxBranches = PartialUnrollingMaxBranches;
178 MaxOps = PartialUnrollingThreshold;
179 } else if (ST->isAtom()) {
180 // On the Atom, the throughput for taken branches is 2 cycles. For small
181 // simple loops, expand by a small factor to hide the backedge cost.
184 } else if (ST->hasFSGSBase() && ST->hasXOP() /* Steamroller and later */) {
187 } else if (ST->hasFMA4() /* Any other recent AMD */) {
189 } else if (ST->hasAVX() || ST->hasSSE42() /* Nehalem and later */) {
192 } else if (ST->hasSSSE3() /* Intel Core */) {
199 // Scan the loop: don't unroll loops with calls, and count the potential
200 // number of taken branches (this is somewhat conservative because we're
201 // counting all block transitions as potential branches while in reality some
202 // of these will become implicit via block placement).
203 unsigned MaxDepth = 0;
204 for (df_iterator<BasicBlock*> DI = df_begin(L->getHeader()),
205 DE = df_end(L->getHeader()); DI != DE;) {
206 if (!L->contains(*DI)) {
211 MaxDepth = std::max(MaxDepth, DI.getPathLength());
212 if (MaxDepth > MaxBranches)
215 for (BasicBlock::iterator I = DI->begin(), IE = DI->end(); I != IE; ++I)
216 if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
217 ImmutableCallSite CS(I);
218 if (const Function *F = CS.getCalledFunction()) {
219 if (!isLoweredToCall(F))
229 // Enable runtime and partial unrolling up to the specified size.
230 UP.Partial = UP.Runtime = true;
231 UP.PartialThreshold = UP.PartialOptSizeThreshold = MaxOps;
233 // Set the maximum count based on the loop depth. The maximum number of
234 // branches taken in a loop (including the backedge) is equal to the maximum
235 // loop depth (the DFS path length from the loop header to any block in the
236 // loop). When the loop is unrolled, this depth (except for the backedge
237 // itself) is multiplied by the unrolling factor. This new unrolled depth
238 // must be less than the target-specific maximum branch count (which limits
239 // the number of taken branches in the uop buffer).
241 UP.MaxCount = (MaxBranches-1)/(MaxDepth-1);
244 unsigned X86TTI::getNumberOfRegisters(bool Vector) const {
245 if (Vector && !ST->hasSSE1())
253 unsigned X86TTI::getRegisterBitWidth(bool Vector) const {
255 if (ST->hasAVX()) return 256;
256 if (ST->hasSSE1()) return 128;
266 unsigned X86TTI::getMaximumUnrollFactor() const {
270 // Sandybridge and Haswell have multiple execution ports and pipelined
278 unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
279 OperandValueKind Op1Info,
280 OperandValueKind Op2Info) const {
281 // Legalize the type.
282 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
284 int ISD = TLI->InstructionOpcodeToISD(Opcode);
285 assert(ISD && "Invalid opcode");
287 static const CostTblEntry<MVT::SimpleValueType>
288 AVX2UniformConstCostTable[] = {
289 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence
290 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence
291 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence
292 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence
295 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
297 int Idx = CostTableLookup(AVX2UniformConstCostTable, ISD, LT.second);
299 return LT.first * AVX2UniformConstCostTable[Idx].Cost;
302 static const CostTblEntry<MVT::SimpleValueType> AVX2CostTable[] = {
303 // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
304 // customize them to detect the cases where shift amount is a scalar one.
305 { ISD::SHL, MVT::v4i32, 1 },
306 { ISD::SRL, MVT::v4i32, 1 },
307 { ISD::SRA, MVT::v4i32, 1 },
308 { ISD::SHL, MVT::v8i32, 1 },
309 { ISD::SRL, MVT::v8i32, 1 },
310 { ISD::SRA, MVT::v8i32, 1 },
311 { ISD::SHL, MVT::v2i64, 1 },
312 { ISD::SRL, MVT::v2i64, 1 },
313 { ISD::SHL, MVT::v4i64, 1 },
314 { ISD::SRL, MVT::v4i64, 1 },
316 { ISD::SHL, MVT::v32i8, 42 }, // cmpeqb sequence.
317 { ISD::SHL, MVT::v16i16, 16*10 }, // Scalarized.
319 { ISD::SRL, MVT::v32i8, 32*10 }, // Scalarized.
320 { ISD::SRL, MVT::v16i16, 8*10 }, // Scalarized.
322 { ISD::SRA, MVT::v32i8, 32*10 }, // Scalarized.
323 { ISD::SRA, MVT::v16i16, 16*10 }, // Scalarized.
324 { ISD::SRA, MVT::v4i64, 4*10 }, // Scalarized.
326 // Vectorizing division is a bad idea. See the SSE2 table for more comments.
327 { ISD::SDIV, MVT::v32i8, 32*20 },
328 { ISD::SDIV, MVT::v16i16, 16*20 },
329 { ISD::SDIV, MVT::v8i32, 8*20 },
330 { ISD::SDIV, MVT::v4i64, 4*20 },
331 { ISD::UDIV, MVT::v32i8, 32*20 },
332 { ISD::UDIV, MVT::v16i16, 16*20 },
333 { ISD::UDIV, MVT::v8i32, 8*20 },
334 { ISD::UDIV, MVT::v4i64, 4*20 },
337 // Look for AVX2 lowering tricks.
339 if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
340 (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
341 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
342 // On AVX2, a packed v16i16 shift left by a constant build_vector
343 // is lowered into a vector multiply (vpmullw).
346 int Idx = CostTableLookup(AVX2CostTable, ISD, LT.second);
348 return LT.first * AVX2CostTable[Idx].Cost;
351 static const CostTblEntry<MVT::SimpleValueType>
352 SSE2UniformConstCostTable[] = {
353 // We don't correctly identify costs of casts because they are marked as
355 // Constant splats are cheaper for the following instructions.
356 { ISD::SHL, MVT::v16i8, 1 }, // psllw.
357 { ISD::SHL, MVT::v8i16, 1 }, // psllw.
358 { ISD::SHL, MVT::v4i32, 1 }, // pslld
359 { ISD::SHL, MVT::v2i64, 1 }, // psllq.
361 { ISD::SRL, MVT::v16i8, 1 }, // psrlw.
362 { ISD::SRL, MVT::v8i16, 1 }, // psrlw.
363 { ISD::SRL, MVT::v4i32, 1 }, // psrld.
364 { ISD::SRL, MVT::v2i64, 1 }, // psrlq.
366 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb.
367 { ISD::SRA, MVT::v8i16, 1 }, // psraw.
368 { ISD::SRA, MVT::v4i32, 1 }, // psrad.
370 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence
371 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence
372 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence
375 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
377 int Idx = CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second);
379 return LT.first * SSE2UniformConstCostTable[Idx].Cost;
382 if (ISD == ISD::SHL &&
383 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
385 if ((VT == MVT::v8i16 && ST->hasSSE2()) ||
386 (VT == MVT::v4i32 && ST->hasSSE41()))
387 // Vector shift left by non uniform constant can be lowered
388 // into vector multiply (pmullw/pmulld).
390 if (VT == MVT::v4i32 && ST->hasSSE2())
391 // A vector shift left by non uniform constant is converted
392 // into a vector multiply; the new multiply is eventually
393 // lowered into a sequence of shuffles and 2 x pmuludq.
397 static const CostTblEntry<MVT::SimpleValueType> SSE2CostTable[] = {
398 // We don't correctly identify costs of casts because they are marked as
400 // For some cases, where the shift amount is a scalar we would be able
401 // to generate better code. Unfortunately, when this is the case the value
402 // (the splat) will get hoisted out of the loop, thereby making it invisible
403 // to ISel. The cost model must return worst case assumptions because it is
404 // used for vectorization and we don't want to make vectorized code worse
406 { ISD::SHL, MVT::v16i8, 30 }, // cmpeqb sequence.
407 { ISD::SHL, MVT::v8i16, 8*10 }, // Scalarized.
408 { ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul.
409 { ISD::SHL, MVT::v2i64, 2*10 }, // Scalarized.
410 { ISD::SHL, MVT::v4i64, 4*10 }, // Scalarized.
412 { ISD::SRL, MVT::v16i8, 16*10 }, // Scalarized.
413 { ISD::SRL, MVT::v8i16, 8*10 }, // Scalarized.
414 { ISD::SRL, MVT::v4i32, 4*10 }, // Scalarized.
415 { ISD::SRL, MVT::v2i64, 2*10 }, // Scalarized.
417 { ISD::SRA, MVT::v16i8, 16*10 }, // Scalarized.
418 { ISD::SRA, MVT::v8i16, 8*10 }, // Scalarized.
419 { ISD::SRA, MVT::v4i32, 4*10 }, // Scalarized.
420 { ISD::SRA, MVT::v2i64, 2*10 }, // Scalarized.
422 // It is not a good idea to vectorize division. We have to scalarize it and
423 // in the process we will often end up having to spilling regular
424 // registers. The overhead of division is going to dominate most kernels
425 // anyways so try hard to prevent vectorization of division - it is
426 // generally a bad idea. Assume somewhat arbitrarily that we have to be able
427 // to hide "20 cycles" for each lane.
428 { ISD::SDIV, MVT::v16i8, 16*20 },
429 { ISD::SDIV, MVT::v8i16, 8*20 },
430 { ISD::SDIV, MVT::v4i32, 4*20 },
431 { ISD::SDIV, MVT::v2i64, 2*20 },
432 { ISD::UDIV, MVT::v16i8, 16*20 },
433 { ISD::UDIV, MVT::v8i16, 8*20 },
434 { ISD::UDIV, MVT::v4i32, 4*20 },
435 { ISD::UDIV, MVT::v2i64, 2*20 },
439 int Idx = CostTableLookup(SSE2CostTable, ISD, LT.second);
441 return LT.first * SSE2CostTable[Idx].Cost;
444 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTable[] = {
445 // We don't have to scalarize unsupported ops. We can issue two half-sized
446 // operations and we only need to extract the upper YMM half.
447 // Two ops + 1 extract + 1 insert = 4.
448 { ISD::MUL, MVT::v16i16, 4 },
449 { ISD::MUL, MVT::v8i32, 4 },
450 { ISD::SUB, MVT::v8i32, 4 },
451 { ISD::ADD, MVT::v8i32, 4 },
452 { ISD::SUB, MVT::v4i64, 4 },
453 { ISD::ADD, MVT::v4i64, 4 },
454 // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
455 // are lowered as a series of long multiplies(3), shifts(4) and adds(2)
456 // Because we believe v4i64 to be a legal type, we must also include the
457 // split factor of two in the cost table. Therefore, the cost here is 18
459 { ISD::MUL, MVT::v4i64, 18 },
462 // Look for AVX1 lowering tricks.
463 if (ST->hasAVX() && !ST->hasAVX2()) {
466 // v16i16 and v8i32 shifts by non-uniform constants are lowered into a
467 // sequence of extract + two vector multiply + insert.
468 if (ISD == ISD::SHL && (VT == MVT::v8i32 || VT == MVT::v16i16) &&
469 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)
472 int Idx = CostTableLookup(AVX1CostTable, ISD, VT);
474 return LT.first * AVX1CostTable[Idx].Cost;
477 // Custom lowering of vectors.
478 static const CostTblEntry<MVT::SimpleValueType> CustomLowered[] = {
479 // A v2i64/v4i64 and multiply is custom lowered as a series of long
480 // multiplies(3), shifts(4) and adds(2).
481 { ISD::MUL, MVT::v2i64, 9 },
482 { ISD::MUL, MVT::v4i64, 9 },
484 int Idx = CostTableLookup(CustomLowered, ISD, LT.second);
486 return LT.first * CustomLowered[Idx].Cost;
488 // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle,
489 // 2x pmuludq, 2x shuffle.
490 if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() &&
494 // Fallback to the default implementation.
495 return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info,
499 unsigned X86TTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
501 // We only estimate the cost of reverse shuffles.
502 if (Kind != SK_Reverse)
503 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
505 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
507 if (LT.second.getSizeInBits() > 128)
508 Cost = 3; // Extract + insert + copy.
510 // Multiple by the number of parts.
511 return Cost * LT.first;
514 unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
515 int ISD = TLI->InstructionOpcodeToISD(Opcode);
516 assert(ISD && "Invalid opcode");
518 std::pair<unsigned, MVT> LTSrc = TLI->getTypeLegalizationCost(Src);
519 std::pair<unsigned, MVT> LTDest = TLI->getTypeLegalizationCost(Dst);
521 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
523 // These are somewhat magic numbers justified by looking at the output of
524 // Intel's IACA, running some kernels and making sure when we take
525 // legalization into account the throughput will be overestimated.
526 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
527 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
528 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
529 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
530 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
531 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
532 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
533 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
534 // There are faster sequences for float conversions.
535 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
536 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
537 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
538 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
539 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
540 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
541 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
542 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
545 if (ST->hasSSE2() && !ST->hasAVX()) {
547 ConvertCostTableLookup(SSE2ConvTbl, ISD, LTDest.second, LTSrc.second);
549 return LTSrc.first * SSE2ConvTbl[Idx].Cost;
552 EVT SrcTy = TLI->getValueType(Src);
553 EVT DstTy = TLI->getValueType(Dst);
555 // The function getSimpleVT only handles simple value types.
556 if (!SrcTy.isSimple() || !DstTy.isSimple())
557 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
559 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
560 AVX2ConversionTbl[] = {
561 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
562 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
563 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
564 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
565 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
566 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
567 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
568 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
569 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
570 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
571 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 3 },
572 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 3 },
573 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
574 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
575 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
576 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
578 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 },
579 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 },
580 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 },
581 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 },
582 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 },
583 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 4 },
586 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
587 AVXConversionTbl[] = {
588 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
589 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
590 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 },
591 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 },
592 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 7 },
593 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 4 },
594 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 },
595 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 4 },
596 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 },
597 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 },
598 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 6 },
599 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 },
600 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 6 },
601 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
602 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 4 },
603 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 4 },
605 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 4 },
606 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 4 },
607 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 4 },
608 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 },
609 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 },
610 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 4 },
611 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 9 },
613 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 },
614 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 8 },
615 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
616 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 },
617 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
618 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
619 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 3 },
620 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
621 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 },
622 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i8, 3 },
623 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i16, 3 },
624 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 },
626 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 },
627 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 5 },
628 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
629 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 9 },
630 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 },
631 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 2 },
632 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
633 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 6 },
634 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 },
635 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 },
636 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 },
637 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 },
638 // The generic code to compute the scalar overhead is currently broken.
639 // Workaround this limitation by estimating the scalarization overhead
640 // here. We have roughly 10 instructions per scalar element.
641 // Multiply that by the vector width.
642 // FIXME: remove that when PR19268 is fixed.
643 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
644 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 4*10 },
646 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 7 },
647 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 },
648 // This node is expanded into scalarized operations but BasicTTI is overly
649 // optimistic estimating its cost. It computes 3 per element (one
650 // vector-extract, one scalar conversion and one vector-insert). The
651 // problem is that the inserts form a read-modify-write chain so latency
652 // should be factored in too. Inflating the cost per element by 1.
653 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 8*4 },
654 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4*4 },
658 int Idx = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
659 DstTy.getSimpleVT(), SrcTy.getSimpleVT());
661 return AVX2ConversionTbl[Idx].Cost;
665 int Idx = ConvertCostTableLookup(AVXConversionTbl, ISD, DstTy.getSimpleVT(),
666 SrcTy.getSimpleVT());
668 return AVXConversionTbl[Idx].Cost;
671 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
674 unsigned X86TTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
675 Type *CondTy) const {
676 // Legalize the type.
677 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
681 int ISD = TLI->InstructionOpcodeToISD(Opcode);
682 assert(ISD && "Invalid opcode");
684 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTbl[] = {
685 { ISD::SETCC, MVT::v2f64, 1 },
686 { ISD::SETCC, MVT::v4f32, 1 },
687 { ISD::SETCC, MVT::v2i64, 1 },
688 { ISD::SETCC, MVT::v4i32, 1 },
689 { ISD::SETCC, MVT::v8i16, 1 },
690 { ISD::SETCC, MVT::v16i8, 1 },
693 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTbl[] = {
694 { ISD::SETCC, MVT::v4f64, 1 },
695 { ISD::SETCC, MVT::v8f32, 1 },
696 // AVX1 does not support 8-wide integer compare.
697 { ISD::SETCC, MVT::v4i64, 4 },
698 { ISD::SETCC, MVT::v8i32, 4 },
699 { ISD::SETCC, MVT::v16i16, 4 },
700 { ISD::SETCC, MVT::v32i8, 4 },
703 static const CostTblEntry<MVT::SimpleValueType> AVX2CostTbl[] = {
704 { ISD::SETCC, MVT::v4i64, 1 },
705 { ISD::SETCC, MVT::v8i32, 1 },
706 { ISD::SETCC, MVT::v16i16, 1 },
707 { ISD::SETCC, MVT::v32i8, 1 },
711 int Idx = CostTableLookup(AVX2CostTbl, ISD, MTy);
713 return LT.first * AVX2CostTbl[Idx].Cost;
717 int Idx = CostTableLookup(AVX1CostTbl, ISD, MTy);
719 return LT.first * AVX1CostTbl[Idx].Cost;
722 if (ST->hasSSE42()) {
723 int Idx = CostTableLookup(SSE42CostTbl, ISD, MTy);
725 return LT.first * SSE42CostTbl[Idx].Cost;
728 return TargetTransformInfo::getCmpSelInstrCost(Opcode, ValTy, CondTy);
731 unsigned X86TTI::getVectorInstrCost(unsigned Opcode, Type *Val,
732 unsigned Index) const {
733 assert(Val->isVectorTy() && "This must be a vector type");
736 // Legalize the type.
737 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Val);
739 // This type is legalized to a scalar type.
740 if (!LT.second.isVector())
743 // The type may be split. Normalize the index to the new type.
744 unsigned Width = LT.second.getVectorNumElements();
745 Index = Index % Width;
747 // Floating point scalars are already located in index #0.
748 if (Val->getScalarType()->isFloatingPointTy() && Index == 0)
752 return TargetTransformInfo::getVectorInstrCost(Opcode, Val, Index);
755 unsigned X86TTI::getScalarizationOverhead(Type *Ty, bool Insert,
756 bool Extract) const {
757 assert (Ty->isVectorTy() && "Can only scalarize vectors");
760 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
762 Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
764 Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
770 unsigned X86TTI::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
771 unsigned AddressSpace) const {
772 // Handle non-power-of-two vectors such as <3 x float>
773 if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
774 unsigned NumElem = VTy->getVectorNumElements();
776 // Handle a few common cases:
778 if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
779 // Cost = 64 bit store + extract + 32 bit store.
783 if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
784 // Cost = 128 bit store + unpack + 64 bit store.
787 // Assume that all other non-power-of-two numbers are scalarized.
788 if (!isPowerOf2_32(NumElem)) {
789 unsigned Cost = TargetTransformInfo::getMemoryOpCost(Opcode,
790 VTy->getScalarType(),
793 unsigned SplitCost = getScalarizationOverhead(Src,
794 Opcode == Instruction::Load,
795 Opcode==Instruction::Store);
796 return NumElem * Cost + SplitCost;
800 // Legalize the type.
801 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
802 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
805 // Each load/store unit costs 1.
806 unsigned Cost = LT.first * 1;
808 // On Sandybridge 256bit load/stores are double pumped
809 // (but not on Haswell).
810 if (LT.second.getSizeInBits() > 128 && !ST->hasAVX2())
816 unsigned X86TTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
817 // Address computations in vectorized code with non-consecutive addresses will
818 // likely result in more instructions compared to scalar code where the
819 // computation can more often be merged into the index mode. The resulting
820 // extra micro-ops can significantly decrease throughput.
821 unsigned NumVectorInstToHideOverhead = 10;
823 if (Ty->isVectorTy() && IsComplex)
824 return NumVectorInstToHideOverhead;
826 return TargetTransformInfo::getAddressComputationCost(Ty, IsComplex);
829 unsigned X86TTI::getReductionCost(unsigned Opcode, Type *ValTy,
830 bool IsPairwise) const {
832 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
836 int ISD = TLI->InstructionOpcodeToISD(Opcode);
837 assert(ISD && "Invalid opcode");
839 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
840 // and make it as the cost.
842 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblPairWise[] = {
843 { ISD::FADD, MVT::v2f64, 2 },
844 { ISD::FADD, MVT::v4f32, 4 },
845 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
846 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
847 { ISD::ADD, MVT::v8i16, 5 },
850 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblPairWise[] = {
851 { ISD::FADD, MVT::v4f32, 4 },
852 { ISD::FADD, MVT::v4f64, 5 },
853 { ISD::FADD, MVT::v8f32, 7 },
854 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
855 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
856 { ISD::ADD, MVT::v4i64, 5 }, // The data reported by the IACA tool is "4.8".
857 { ISD::ADD, MVT::v8i16, 5 },
858 { ISD::ADD, MVT::v8i32, 5 },
861 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblNoPairWise[] = {
862 { ISD::FADD, MVT::v2f64, 2 },
863 { ISD::FADD, MVT::v4f32, 4 },
864 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
865 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3".
866 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3".
869 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblNoPairWise[] = {
870 { ISD::FADD, MVT::v4f32, 3 },
871 { ISD::FADD, MVT::v4f64, 3 },
872 { ISD::FADD, MVT::v8f32, 4 },
873 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
874 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "2.8".
875 { ISD::ADD, MVT::v4i64, 3 },
876 { ISD::ADD, MVT::v8i16, 4 },
877 { ISD::ADD, MVT::v8i32, 5 },
882 int Idx = CostTableLookup(AVX1CostTblPairWise, ISD, MTy);
884 return LT.first * AVX1CostTblPairWise[Idx].Cost;
887 if (ST->hasSSE42()) {
888 int Idx = CostTableLookup(SSE42CostTblPairWise, ISD, MTy);
890 return LT.first * SSE42CostTblPairWise[Idx].Cost;
894 int Idx = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy);
896 return LT.first * AVX1CostTblNoPairWise[Idx].Cost;
899 if (ST->hasSSE42()) {
900 int Idx = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy);
902 return LT.first * SSE42CostTblNoPairWise[Idx].Cost;
906 return TargetTransformInfo::getReductionCost(Opcode, ValTy, IsPairwise);
909 unsigned X86TTI::getIntImmCost(const APInt &Imm, Type *Ty) const {
910 assert(Ty->isIntegerTy());
912 unsigned BitSize = Ty->getPrimitiveSizeInBits();
919 if (Imm.getBitWidth() <= 64 &&
920 (isInt<32>(Imm.getSExtValue()) || isUInt<32>(Imm.getZExtValue())))
923 return 2 * TCC_Basic;
926 unsigned X86TTI::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
928 assert(Ty->isIntegerTy());
930 unsigned BitSize = Ty->getPrimitiveSizeInBits();
934 unsigned ImmIdx = ~0U;
936 default: return TCC_Free;
937 case Instruction::GetElementPtr:
938 // Always hoist the base address of a GetElementPtr. This prevents the
939 // creation of new constants for every base constant that gets constant
940 // folded with the offset.
942 return 2 * TCC_Basic;
944 case Instruction::Store:
947 case Instruction::Add:
948 case Instruction::Sub:
949 case Instruction::Mul:
950 case Instruction::UDiv:
951 case Instruction::SDiv:
952 case Instruction::URem:
953 case Instruction::SRem:
954 case Instruction::Shl:
955 case Instruction::LShr:
956 case Instruction::AShr:
957 case Instruction::And:
958 case Instruction::Or:
959 case Instruction::Xor:
960 case Instruction::ICmp:
963 case Instruction::Trunc:
964 case Instruction::ZExt:
965 case Instruction::SExt:
966 case Instruction::IntToPtr:
967 case Instruction::PtrToInt:
968 case Instruction::BitCast:
969 case Instruction::PHI:
970 case Instruction::Call:
971 case Instruction::Select:
972 case Instruction::Ret:
973 case Instruction::Load:
977 if ((Idx == ImmIdx) &&
978 Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
981 return X86TTI::getIntImmCost(Imm, Ty);
984 unsigned X86TTI::getIntImmCost(Intrinsic::ID IID, unsigned Idx,
985 const APInt &Imm, Type *Ty) const {
986 assert(Ty->isIntegerTy());
988 unsigned BitSize = Ty->getPrimitiveSizeInBits();
993 default: return TCC_Free;
994 case Intrinsic::sadd_with_overflow:
995 case Intrinsic::uadd_with_overflow:
996 case Intrinsic::ssub_with_overflow:
997 case Intrinsic::usub_with_overflow:
998 case Intrinsic::smul_with_overflow:
999 case Intrinsic::umul_with_overflow:
1000 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
1003 case Intrinsic::experimental_stackmap:
1004 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1007 case Intrinsic::experimental_patchpoint_void:
1008 case Intrinsic::experimental_patchpoint_i64:
1009 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
1013 return X86TTI::getIntImmCost(Imm, Ty);